Approximate spectral clustering

Abstract

While spectral clustering has recently shown great promise, computational cost makes it infeasible for use with large data sets. To address this computational challenge, this paper considers the problem of approximate spectral clustering, which enables both the feasibility (of approximately clustering in very large and unloadable data sets) and acceleration (of clustering in loadable data sets), while maintaining acceptable accuracy. We examine and propose several schemes for approximate spectral grouping, and make an empirical comparison of those schemes in combination with several sampling strategies. Experimental results on several synthetic and real-world data sets show that approximate spectral clustering can achieve both the goals of feasibility and acceleration. © Springer-Verlag Berlin Heidelberg 2009.